Quoting Nico Boehr (2023-03-24 15:54:23)
The KVM_S390_GET_CMMA_BITS ioctl may return incorrect values when userspace specifies a start_gfn outside of memslots.
This can occur when a VM has multiple memslots with a hole in between:
+-----+----------+--------+--------+ | ... | Slot N-1 | <hole> | Slot N | +-----+----------+--------+--------+ ^ ^ ^ ^ | | | | GFN A A+B | | A+B+C | A+B+C+D
When userspace specifies a GFN in [A+B, A+B+C), it would expect to get the CMMA values of the first dirty page in Slot N. However, userspace may get a start_gfn of A+B+C+D with a count of 0, hence completely skipping over any dirty pages in slot N.
The error is in kvm_s390_next_dirty_cmma(), which assumes gfn_to_memslot_approx() will return the memslot _below_ the specified GFN when the specified GFN lies outside a memslot. In reality it may return either the memslot below or above the specified GFN.
When a memslot above the specified GFN is returned this happens:
- ofs is calculated, but since the memslot's base_gfn is larger than the specified cur_gfn, ofs will underflow to a huge number.
- ofs is passed to find_next_bit(). Since ofs will exceed the memslot's number of pages, the number of pages in the memslot is returned, completely skipping over all bits in the memslot userspace would be interested in.
Fix this by resetting ofs to zero when a memslot _above_ cur_gfn is returned (cur_gfn < ms->base_gfn).
Signed-off-by: Nico Boehr nrb@linux.ibm.com Reviewed-by: Claudio Imbrenda imbrenda@linux.ibm.com
Since Janosch asked for it, I would suggest the following Fixes tag:
Fixes: afdad61615cc ("KVM: s390: Fix storage attributes migration with memory slots")